Is it reliable to use DeepSeek for scientific research? Researchers: It is equivalent to a PhD level, and it is no problem to build frameworks and write code

Looking around the world, the hottest and most cutting-edge topic nowadays is none other than AI (artificial intelligence). The DeepSeek-R1 large language model, "born" at Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., has astonished the industry as soon as it was unveiled. It became popular during the Spring Festival and set off a huge wave of AI storms. The well-known academic journal Nature recently published an article stating that in terms of solving scientific problems, models such as DeepSeek-R1 have demonstrated the ability to surpass early language models and have the potential for scientific research applications.

DeepSeek makes researchers feel like they have found a treasure? Is it reliable for researchers to use it for research?

How does “doctoral student” DeepSeek conduct scientific research?

"The railroad tracks are cut through by clouds, the road is covered by clouds in autumn, and the old post station has changed its appearance. I recall the journeys of my childhood, the long dragon panting under the moon; the lonely goose covered with frost, carrying a box in the cold window."

These two sentences were posted on the WeChat Moments of Wan Qing, the head of the Functional Materials and Devices Heterogeneous Integration Research Center of Yongjiang Laboratory. With such beautiful sentences and quotations, can you tell that it was written with DeepSeek?

In addition to being a poet, DeepSeek is also a powerful assistant in Wan Qing's work. The topic of brain-like neuromorphic devices that Wan Qing is studying cannot be separated from AI. He has been using AI before, and now he has switched from other AI to DeepSeek. "Its reasoning and language organization abilities are very strong, and the PPT framework it builds is very logical."

Zhang Jiahao, assistant researcher of the computer vision team of the Institute of Technology and Engineering of the Ningbo Institute of Materials, Chinese Academy of Sciences, focuses on the design of robotic equipment. He usually uses DeepSeek to write code. Robots are to be used in many fields such as agriculture and chemical industry. How can they be better integrated with the automation industry? Zhang Jiahao believes that DeepSeek also plays a good role in providing guidance.

Jiang Chao, a researcher and doctoral supervisor at the Institute of Life Sciences of Zhejiang University, directly regards DeepSeek as an "interdisciplinary consultant". "Although it cannot be as strong as us in a certain field and lacks some details, every field it covers can reach the level of a second or third year doctoral student. It is a true all-rounder."

In the view of Gao Pengcheng, a researcher of the Smart Thermal Power Team of the White Horse Lake Laboratory, DeepSeek, as a benchmark for open source large models, has played a significant enabling role in the traditional thermal power industry. The "open source" here does not mean saving money, but means that the source code of the development software can be used and modified by the public. "The thermal power industry has extremely high requirements for data security. The private deployment of DeepSeek on the factory side can ensure the security and controllability of data, which not only improves the intelligence level and efficiency of production operations, but also does not need to worry about data leakage." Gao Pengcheng explained.

Is “answering the wrong question” the secret to “a flash of inspiration” in scientific research?

When encountering a problem that cannot be solved in scientific research, will AI "lie"?

In fact, even AI that seems to be "extremely smart" cannot escape the trouble of "hallucinations". Wan Qing explained: "AI hallucinations refer to AI generating information that seems reasonable but is actually inaccurate or false."

AI hallucinations may have a big impact. In the legal field, it may cite fictitious legal provisions and cases. In the field of journalism, it may generate false news, disrupt the order of information dissemination, and conduct erroneous public opinion guidance. "When using DeepSeek in scientific research before, I found that the generated references seemed to be 'well-formed', but when I actually found the document, I found that it was 'irrelevant' and was just a forced copy." Wan Qing said, "But now if we modify it on this basis and play the role of people, the efficiency will not only be twice the result with half the effort."

AI is like a "double-edged sword". These "out of nothing" discoveries can lead to unexpected discoveries. Wan Qing experienced this in a recent experiment. "I expressed the vague ideas in my mind with five or six keywords and asked DeepSeek questions. It actually generated an article with a paper framework, which gave me a lot of inspiration."

In Jiang Chao's opinion, AI chatbots not only lie, but also rely heavily on the accuracy and depth of questions asked. Everyone can use AI and ask questions to AI, but how to ask questions has become an important topic in the field of scientific research.

"When I ask a broad question like 'What role do microorganisms play in the remediation of contaminated soil?' It only gives traditional answers like degrading microorganisms, biosorption and fixation. But if I ask it more specifically, it comes up with some interesting ideas that may not even have been done before." Jiang Chao believes that asking questions to DeepSeek is a key ability that researchers should have today.

At the forefront of the storm, AI for Science is the general trend

Since its introduction, artificial intelligence has never been at the forefront of human technological progress like it is today. Can AI represented by DeepSeek replace real people and be competent for scientific research positions?

"No!" Gao Pengcheng answered without hesitation. Several other experts unanimously agreed with this question. Although AI technology is developing at an unprecedented speed, generative AI is constantly refreshing people's cognition with its amazing creativity, but its technology is still limited. "Not limited to DeepSeek, AI's knowledge is indeed broad, but it still lacks innovation. Their thinking ability is based on various data theories. When they encounter problems without reference, they cannot answer them." Zhang Jiahao said.

In addition, although AI represented by DeepSeek is powerful, it is still a tool. It is not reliable to rely entirely on it for reading or writing. The results still require people to judge right and wrong, pros and cons, and whether they meet expectations.

AI cannot replace scientific researchers, but it seems that AI has never been so close to scientific research. The 2024 Nobel Prize has pushed AI to the center of the scientific stage. AI is reshaping the world, especially in the paradigm of scientific exploration and innovation, setting off a new wave.

"In the long run, AI for Science (artificial intelligence-driven scientific research, AI4S) is definitely a trend," Gao Pengcheng explained. "AI will be increasingly integrated into work and have a profound impact on scientific research." AI for Science refers to the use of AI technologies and methods to learn, simulate, predict and optimize various phenomena and laws in nature and human society, thereby promoting scientific discovery and innovation. It is believed to bring a new paradigm to scientific research.

Jiang Chao believes that as a researcher, AI can greatly improve research efficiency. From literature search to idea generation, from experimental design to program analysis, AI can help at every step. "With the advancement of AI technology, the more you know, the more it can help you. It is expected that AI will play an increasingly critical role in the field of scientific research in the future," Jiang Chao said.